The anchor-free based face detection methods can cover a large range of scales and perform better in the speed. However, their performance still bears a large gap compared with anchor-based methods, especially for detecting small faces. Because they are troubled by the context modeling and scale imbalance problems. In this study, to address these problems, we propose a novel single shot anchor-free face detector (SAFD) for detecting multi-scale faces by leveraging the multi-scale context aware information of multi-layer features. In the SAFD, we use the dilated convolution layers and attention mechanism to select the informative features that can accommodate to different scales. We also propose a scale-aware sampling strategy to mitigate the scale imbalance problem by adaptivity selecting the positive training samples. The experimental results on two public benchmark datasets, Wider Face and FDDB dataset, demonstrate that our SAFD can achieve competitive performance with the anchor-based detectors while with lower computation cost.

基于无锚的人脸检测方法可以覆盖大范围的尺度，并且速度更快。但是，与基于锚的方法相比，它们的性能仍存在较大差距，尤其是在检测小脸部时。因为它们受到上下文建模和规模不平衡问题的困扰。在这项研究中，为了解决这些问题，我们提出了一种新颖的单发免锚人脸检测器（SAFD），它可以利用多层特征的多尺度上下文感知信息来检测多尺度面孔。在SAFD中，我们使用膨胀的卷积层和注意力机制来选择可以适应不同尺度的信息特征。我们还提出了一种规模感知的抽样策略，通过自适应选择积极的训练样本来缓解规模失衡问题。